Setup fell fromseven hours to ten minutesacross 113 client workspaces.
MTA Group runs eight agencies, and Zero Fluff Digital is one of them. Client context used to live in chat threads, local folders and people’s heads, so every AI session opened with somebody re-explaining the account. This is the system that replaced that ritual, read four months after the first commit, and what it changes for the account you would hand us.
- 5–10 min
- to stand up a new client workspace, down from 3.5–7 hours
- 113
- workspaces on one identical standard, across 8 agencies
- 27
- data sources wired in, with the keys never leaving the gateway
- 28,800
- characters of project context loaded before the first question
Three problems, and none of them was the tooling.
Context that did not survive a holiday
An agency produces context every day: reports, audits, recommendations, what was agreed on a call, why a budget moved. Almost none of it had an address. It lived in a project manager’s head, in a chat thread from three weeks ago, in a spreadsheet on somebody’s drive. When that person went on leave or changed projects, part of the account went with them.
Every session opened on an empty page
An AI tool has exactly the problem a new hire has: it knows nothing. Paste the data, explain the account, recall what was decided last month, mention that this client works in English and that one in Polish. Minutes went by before the first useful question, and the quality of the answer depended on how much anybody felt like pasting that day.
Standing up one project was its own project
Credentials for every data source, connectors to wire and test, an instruction describing the client and the rules, scheduled pulls of transcripts and data, permissions to hand out, then a check that any of it worked. The source puts that at three and a half to seven hours of human time before anybody looked at a number.
Four moves. In this order.
The unit of context became a repository, not a folder
Each client’s context got its own git repository with the same structure as every other one. That buys four things a folder never will: history of who changed what and why, permissions, one source of truth instead of fourteen copies, and a shape that can be enforced from outside. One shared knowledge base for every client was rejected on purpose. One workspace cannot see another’s data, and that boundary is architecture rather than discipline.
Data is fetched through a gateway, never copied in
Two gateways cover 27 sources: 17 for reading, among them Google Ads, Search Console, GA4, Meta Ads, Merchant Center, Bing Webmaster Tools, PageSpeed, call transcripts and the project CRM, and 10 for writing, among them Sheets, Drive, Forms, ClickUp and Firecrawl. The keys stay server-side and never land in a repository. 96 domains get a daily snapshot. A copy of the data would start ageing the minute it was made, and would turn the repository into a place keys can leak from.
The context loads itself, and states what it does not have
A start hook assembles about 28,800 characters before the first question: project identity, authentication status to the data gateway, data freshness with a warning when something has gone stale, the skills available in this specific repository, the map of connected sources, repository state and the policies in force. The section on sources is the one that matters most. It closes the answer “I have no access to Google Ads” when the access is there, and it closes the opposite failure too, because no data for a period reads as no data rather than as an invitation to guess.
One canon, propagated to 113 places
Rules, skills, hooks, templates and guards live in a single parent repository and are pushed down, so a fix written once lands in 113 places in one pass instead of in 113 conversations. The price is discipline: a skill is edited in the canon, never in a copy, because the copy is overwritten on the next propagation. Ten guards then run in every workspace without anyone configuring them, including a client-content linter, protection of the project instruction from overwrite, a warning on a commit from the wrong identity, a secret scan before every push and a quality gate before publication.
Four months after the first commit. Caveats included.
Every figure below comes from a system read taken on 25 August 2026, four months after the first commit on 8 April 2026, and is traceable in the repositories and the telemetry. The chart breaks down where the three and a half to seven hours used to go.
Add the six up and you land on the 3.5 to 7 hours the source reports, and each one now reads zero or runs by itself, because the work happens once centrally instead of once per project. The source is explicit that this is an estimate built from the listed steps rather than a stopwatch measurement, since no baseline was collected before the build started. That is why it stays a range and never becomes a single number.
“Picking up those projects is not always the most pleasant thing, because someone is on holiday, someone is away, someone has no time. With the repositories we can gather people’s work and knowledge in one place. Five minutes, instead of catching the project manager and passing everything on by word of mouth.”Rado Kmita, system architect, MTA Group
What you get from this is a team that starts at minute one.
The number worth caring about is not 113. It is that the people on your account do not spend the first part of a session rebuilding what was agreed last month, and that whoever covers for someone on leave opens the same structure rather than a stranger’s folder. Your data stays where it is: the workspace knows where to reach it and how, and the keys never leave the gateway. Three commitments sit outside the repository and are worth stating plainly. The model provider works under a data processing agreement, client contracts were updated for it, and where AI co-writes material for a client, the client is told and the material is marked.
Two things here cannot be promised. The first is a saving in your day-to-day work. What was measured is the setup, not the hours after it, and the source says so rather than rounding in its own favour. Usage telemetry only started on 5 August 2026: the last thirty days show 270 skill invocations across 25 skills, of which 242 came from a single scheduled automation, leaving roughly 28 real human uses. That is too little to call adoption, and calling it adoption is exactly the move this page exists to avoid.
The second is that none of this makes a decision for you. It removes the setup tax and the re-explaining, and it closes the two failure modes that produce confident wrong answers: a tool that claims it has no access when it has, and a tool that fills a hole in the data with a guess. What to do about a result is still a person’s call. On your account, that person is one of ours, working in the system above rather than in a folder you will never see.
This work was delivered by MTA Group, which Zero Fluff Digital is part of, by the same specialists who would work on your account. Every figure here comes from the case study MTA Group published, republished here with their consent. The reviews behind it are public and verified on Clutch, where clients rate the work rather than the agency describing itself.
Want to know where your account context actually lives?
30 minutes, no deck. You get an honest read on what happens to your account when the person who holds it in their head is away, and what rebuilding that costs you every time.
Book a callNext case studyAudioteka: 7M organic TikTok views in 12 months →